Remote Sen. 2009, 1, 355-374; doi:10.3390/rs1030355 OPEN ACCESS
Remote Sensing ISSN 2072-4292 www.mdpi.com/journal/remotesensing Article
A Simple Algorithm for Large-Scale Mapping of Evergreen Forests in Tropical America, Africa and Asia Xiangming Xiao 1,2,*, Chandrashekhar M. Biradar and Michael Keller 2,4 1
2
3 4
1,2,
*, Christina Czarnecki 2, Tunrayo Alabi
3
Department of Botany and Microbiology, and Center for Spatial Analysis, University of Oklahoma, Norman, OK 73019, USA Institute for the Study of Earth, Oceans and Space, University of New Hampshire, Durham NH 03824, USA; E-Mail:
[email protected] (C.C.) International Institute of Tropical Agriculture, Ibadan, Nigeria; E-Mail:
[email protected] (T.A.) The National Ecological Observatory Network (NEON), Boulder, CO, USA; E-Mail:
[email protected] (M.K.)
* Author to whom correspondence should be addressed; E-Mails:
[email protected] (X.X.);
[email protected] (C.B.); Tel.: +1-405-325-8941; Fax: +1-405-325-3442 Received: 28 April 2009; in revised form: 30 May 2009 / Accepted: 3 August 2009 / Published: 12 August 2009
Abstract: The areal extent and spatial distribution of evergreen forests in the tropical zones are important for the study of climate, carbon cycle and biodiversity. However, frequent cloud cover in the tropical regions makes mapping evergreen forests a challenging task. In this study we developed a simple and novel mapping algorithm that is based on the temporal profile analysis of Land Surface Water Index (LSWI), which is calculated as a normalized ratio between near infrared and shortwave infrared spectral bands. The 8-day composites of MODIS Land Surface Reflectance data (MOD09A1) in 2001 at 500-m spatial resolution were used to calculate LSWI. The LSWI-based mapping algorithm was applied to map evergreen forests in tropical Africa, America and Asia (30°N–30°S). The resultant maps of evergreen forests in the tropical zone in 2001, as estimated by the LSWI-based algorithm, are compared to the three global forest datasets [FAO FRA 2000, GLC2000 and the standard MODIS Land Cover Product (MOD12Q1) produced by the MODIS Land Science Team] that are developed through complex algorithms and processes. The inter-comparison of the four datasets shows that the area estimate of evergreen forest from the LSWI-based algorithm fall within the range of forest area estimates from the FAO FRA 2000, GLC2000 and MOD12Q1 at a country level.
Remote Sens. 2009, 9
356
The area and spatial distribution of evergreen forests from the LSWI-based algorithm is to a large degree similar to those of the MOD12Q1 produced by complex mapping algorithms. The results from this study demonstrate the potential of the LSWI-based mapping algorithm for large-scale mapping of evergreen forests in the tropical zone at moderate spatial resolution. Keywords: MODIS image; land surface water index; temporal profile analysis; evergreen forests
1. Introduction Evergreen forests (both broadleaf and needle leaf trees) in the tropical zone are an essential timber resource and play an important role in the global carbon and water cycles, biodiversity and climate. A number of efforts have been devoted to quantify the areas and spatial distributions of tropical forests [1-6]. However, frequent cloud cover in the tropical regions makes mapping evergreen forests in these zones a challenging task. In general, three research approaches have been widely used to quantify the area and spatial distribution of evergreen tropical forests at local, continental and global scales. One approach is to compile forest inventory statistics at different administrative unit levels (county, province and nation) in a region, for example, the United Nations Food and Agriculture Organization (FAO) produced Global Forest Resources Assessments (GFRAs) in 1990, 2000 and 2005, based on forest statistics provided by individual countries [7,8]. The second approach is to map forests using satellite images at fine spatial resolution (tens of meters), e.g., Landsat TM (Thematic Mapper) and ETM+ (Enhanced Thematic Mapper). Landsat TM/ETM+ images have a spatial resolution of 30-m, and are widely used to map forests and deforestation in Amazon [9], and the globe [10]. Global-scale mapping of evergreen forests in the tropical zone from satellite images at fine resolution (e.g. Landsat TM/ETM+) is extremely challenging, because frequent cloud coverage in the moist tropical zone and infrequent image acquisition (due to the 16-day revisit interval by Landsat) often result in few cloud-free Landsat images available for analysis. Therefore, to generate a wall-to-wall coverage of Landsat TM/ETM+ images for the global tropical zone one usually needs to obtain images from several years of image acquisition by Landsat TM/ETM+ sensors. The third approach is to map forests using satellite images at moderate spatial resolution (hundreds of meters), e.g., Advanced Very High Resolution Radiometer (AVHRR) sensors [3], SPOT-Vegetation (VGT) sensors [11] and Moderate Resolution Imaging Spectroradiometer (MODIS) sensors [4,12]. These moderate-resolution sensors acquire daily images for the globe and provide time series image data for land cover classification. The Global Land Cover Characteristics (GLCC, DIScover dataset) dataset used AVHRR data at 1-km resolution in 1992–1993 [13]. The Global Land Cover 2000 (GLC2000) dataset used the Vegetation data at 1-km resolution in 2000 [6,11,14]. The Global Land Cover Data (MOD12Q1) used MODIS data at 1-km resolution [15]. All these data products were generated from supervised classification algorithms, which require substantial training datasets in the ground and experienced users to interpret and label spectral clusters into individual land cover types. Due to frequent cloud cover and large temporal variation of cloud cover, cloud-free time series image datasets vary significantly between years, which may have substantial impacts on the statistics of spectral clusters and
Remote Sens. 2009, 9
357
interpretation of spectral clusters into land cover types. Although it is possible to apply these complex mapping algorithms to generate annual maps of forests, it is often time consuming and expensive, as it requires updating the training data periodically. To directly overlay two annual forest maps and then calculate annual rates of deforestation in the world is often a challenging task, because different image data sources, training datasets, and statistical algorithms are used. Here we present a study that aims to develop a simple and novel algorithm to map the evergreen forests in the tropical world, using multi-temporal MODIS data in a year. If the simple and novel approach could produce evergreen forest maps that are similar to the forest maps from the above-mentioned complicated mapping algorithms [6,11,15], it may offer the potential for us to generate annual maps of evergreen forests in the near future, which is needed for rapid assessment of forest resources in the world. 2. Satellite Imagery and Mapping Algorithm 2.1. MODIS Land Surface Reflectance Data and Vegetation Indices The MODIS sensor onboard the NASA Terra satellite has 36 spectral bands, and seven of these 36 bands are primarily designed for the study of vegetation and land surface: blue (459–479 nm), green (545–565 nm), red (620–670 nm), near infrared (841–875 nm, 1,230–1,250 nm) and shortwave infrared (1,628–1,652 nm, 2,105–2,155 nm). The red and NIR1 (841–875 nm) bands have a spatial resolution of 250-m, and the other five bands (blue, green, NIR2, SWIR1, SWIR2 bands) have a spatial resolution of 500-m. The MODIS sensor acquires daily imagery for the globe. The MODIS Land Science Team provides a suite of standard MODIS data products to the users, including the 8-day composite MODIS Land Surface Reflectance Product (MOD09A1). There are forty-six 8-day composites in a year. Each 8-day composite (MOD09A1) includes estimates of land surface reflectance for the seven spectral bands at 500-m spatial resolution. In the production of MOD09A1, atmospheric corrections for gases, thin cirrus clouds and aerosols are implemented [16]. MOD09A1 8-day composites are generated in a multi-step process that first eliminates pixels with a low observational coverage, and then selects an observation with highest quality during the 8-day period [17]. The MOD09A1 standard products are organized in a tile system with the Sinusoidal projection; and each tile covers an area of 1,200 × 1,200 km (approximately 10° latitude × 10° longitude at equator). In this study we acquired MOD09A1 data in 2001 (Collection 5) from the USGS EROS Data Center (EDC; http://edc.usgs.gov/); and the MOD09A1 datasets cover the tropical zone (ranging from 30°N to 30°S). For each MOD09A1 file, the quality of individual observations (e.g., clouds, cloud shadow) was identified, and three vegetation indices are calculated: Normalized Difference Vegetation Index (NDVI, Equation 1) [18], Enhanced Vegetation Index (EVI, Equation 2) [19], and Land Surface Water Index (LSWI, Equation 3) [20], using Blue, Red, NIR1 (841–875 nm) and SWIR2 (1,628–1,652 nm) spectral bands. The vegetation indices data products are available to the public (http://www.eomf.ou.edu). NDVI =
ρ nir − ρ red ρ nir + ρ red
(1)
Remote Sens. 2009, 9
358
EVI = 2.5 ×
ρ nir
ρ nir − ρ red + 6 × ρ red − 7.5 × ρ blue + 1
LSWI =
ρ nir − ρ swir ρ nir + ρ swir
(2)
(3)
The shortwave infrared (SWIR) spectral band is sensitive to vegetation water content and soil moisture [21], and a combination of NIR and SWIR bands have been used to derive water sensitive vegetation indices [22-27], including Land Surface Water Index (LSWI). LSWI is sensitive to equivalent water thickness (EWT, g H2O/m2) [27-29]. And recently LSWI has been used for mapping forests and agriculture [30,31], inundation [30,32], vegetation phenology [33,34], and gross primary production of forests [35]. 2.2. Temporal Profile Analysis for Identifying and Mapping Evergreen Forests A green leaf has higher NIR reflectance than SWIR reflectance, resulting in a LSWI value of above 0.0 (positive value). A senescent leaf and soil have lower NIR reflectance than SWIR reflectance, resulting in a LSWI value of below 0.0 (negative value). Spectral reflectance of plants and soils are well documented and reported in many hyperspectral libraries, for example, the spectral libraries in the USGS Spectroscopy Lab (http://speclab.cr.usgs.gov/) and the commercial ENVI image processing software (http://www.ittvis.com/ProductServices/ENVI.aspx). For plant leaves, LSWI > 0 or LSWI < 0 represents a state of change from green leaf to senescent leaf, a phenology (leaf aging process)-related change in biophysical property of leaf. In an early study the seasonal dynamics of three vegetation indices (LSWI, NDVI and EVI) were examined for seven forest types (four deciduous broadleaf forests, one deciduous needle leaf forest, two mixed forests and one evergreen needle leaf forest) in Northeastern China [26]. LSWI values of evergreen needle leaf forest remain >0.0 for all good-quality satellite observations throughout a year, while all the other six forest types have some observations with LSWI values of 0.0 for all cloud-free observations [29]. Tropical regions have a variety of land cover types, as an example, Figure 1 shows the seasonal dynamics of LSWI of individual pixels from six land cover types (evergreen broadleaf forest, deciduous broadleaf forest, shrubland, cropland, grassland and desert) in tropical Africa. All LSWI values of the desert pixel in a year are below - 0.1, and have little seasonal variation in a year. LSWI values of evergreen broadleaf forest remain >0.0 for all good-quality observations throughout a year, while all the other five land cover types have a number of observations with LSWI < 0.0 values in a year (Figure 1). Another previous study for inundated paddy rice fields, one of wetlands, have shown that paddy rice fields have a number of observations with LSWI
